OpenAI will begin adding invisible watermarks to text generated by ChatGPT and Codex for users in the European Union, bringing the company’s text-generation products under a new layer of AI-content transparency requirements introduced by the EU AI Act.
The company said Monday that the watermarking system will be rolled out over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU. OpenAI will not initially make text watermarking a global default.
Developers using OpenAI’s API outside the European Union will also be able to activate the technology for selected models from Monday, although it will remain switched off by default.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
The move marks a significant shift in how AI-generated text can be identified. Unlike a visible label or symbol, OpenAI’s system embeds a statistical pattern directly into the generated words, allowing a detector to assess whether text was produced or processed by an OpenAI model.
The change comes as regulators and AI companies grapple with a difficult question: how can synthetic content be identified without making AI-generated material visibly different from human writing?
Under the EU AI Act’s transparency provisions, which took effect on August 2, providers of AI systems are required to ensure that certain AI-generated or manipulated content can be detected by appropriate technical means.
OpenAI’s approach attempts to meet that requirement without interrupting the way users read or share AI-generated text.
Watermark Is Embedded In Word Choices
The system does not insert a special character, image, or visible mark into a document.
Instead, OpenAI subtly influences the model’s selection of words as it generates text. The changes are designed to be imperceptible to readers but sufficiently structured for a detector with access to the relevant detection mechanism to identify the statistical signature. Because the watermark is embedded in the text itself, it can remain attached to the content when users copy and paste it into another application.
OpenAI has named the underlying technique textGrain and published a technical report describing how it works. The report was co-written with researchers from the University of Pennsylvania and Yale.
At a simplified level, the method uses a secret key to organize possible next-word predictions during generation. By repeatedly applying small changes to those selections, the system produces a statistical pattern that becomes increasingly detectable as more text is generated.
That yields a fundamentally different approach from visible AI labels. A reader does not need to see or preserve a tag for the signal to travel with the text. OpenAI said it found no meaningful change in model performance when the watermark was enabled. But the technology is not designed to establish definitive authorship.
OpenAI’s own testing found that relatively modest editing could substantially weaken detection. In one test, replacing 10% of words with synonyms reduced detection accuracy from about 92% to 66%.
Short passages also present difficulties, while mathematical answers and translated text are harder to identify reliably. Those limitations mean the system is better understood as an indicator than as a forensic proof of AI authorship.
OpenAI explicitly cautioned that the absence of a watermark does not demonstrate that a person wrote the material. Text could be too short, heavily edited, or generated by another AI system that does not use the same watermarking method.
“[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” the company said.
The clarification matters in education, publishing and professional settings where AI-assisted writing increasingly combines human decisions with machine-generated material.
Detection Raises Questions About Human-AI Collaboration
OpenAI will initially restrict access to its text watermark detector to approved researchers and expert organizations.
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” the company said.
The restricted rollout reflects the uncertainty surrounding AI detection generally. A detector that incorrectly labels human writing as AI-generated could have serious consequences, while a system that misses AI-generated text could provide false confidence.
The problem becomes even more complicated when people substantially edit AI-generated material.
A piece of writing may begin with ChatGPT, undergo several rounds of human editing, and eventually contain a mixture of machine-generated and human-authored material. A watermark can potentially establish that an OpenAI system contributed to the text, but it cannot determine the proportion of human input.
That limitation also distinguishes content provenance from authorship. The technology is intended to establish a signal about the origin or processing of content, not settle questions about who ultimately created it.
OpenAI’s decision also comes amid growing pressure on AI companies to make generated content identifiable.
Anthropic said two months ago that it would watermark text generated by Claude globally. The decision prompted criticism from some Claude users, who argued that users provide the “instructions, context, decisions” while the model functions as a tool.
OpenAI had previously developed a text watermarking system but decided not to release it, partly because of concerns that users could migrate to competing AI systems that did not watermark their output, according to a Wall Street Journal report in 2024.
The European Union has now provided a regulatory incentive for companies to move in the opposite direction.
OpenAI, Anthropic, Google, Meta and Microsoft are among the major AI companies that have committed to following the EU’s code of practice covering AI-generated content.
For OpenAI, limiting mandatory text watermarking to the EU initially reflects the different regulatory environments facing AI companies around the world. Developers using the API can voluntarily activate the technology for selected models globally, but OpenAI is not imposing it as a worldwide default.
The approach also leaves open a broader question for the industry: whether invisible provenance technology will eventually become a standard feature of generative AI or remain primarily a regulatory requirement in jurisdictions such as the EU.
OpenAI’s rollout currently indicates that AI transparency is moving beyond visible labels and disclosure statements toward technical systems embedded directly into the generation process. But the limitations of textGrain also show why watermarking is unlikely to provide a definitive answer to the question of whether a piece of writing was created by a human. It can provide a signal. But it cannot, by itself, reconstruct the creative process behind the words.